Healing Fields: Subject-Specific Pseudo-Healthy Deformations as Endophenotypes for Alzheimer's Genetics

Genetic studies of Alzheimer's disease gain a lot of power when they replace the yes/no diagnosis with continuous brain measurements, such as regional volumes or amyloid levels. Shen and colleagues report that such quantitative traits carry "4 to 10 times more statistical power than traditional binary clinical diagnoses" (Genetic analysis of quantitative phenotypes in AD and MCI). Most of these brain measurements still mix two things. One part is disease-related change. The other is the large, heritable variation in how normal brains are shaped. A small hippocampus may reflect atrophy, or it may simply be that person's healthy anatomy. Genetic variants that affect normal brain shape then crowd the association results and dilute the disease signal.

We propose to separate the two using pseudo-healthy reconstruction, a method from anomaly detection. A generative model trained only on healthy brains finds, for each scan, the closest healthy-looking version of that same brain. The model shows the change as a smooth spatial warp, or deformation, that turns the observed brain into its healthy counterpart. We call this voxel-level map the healing field. It says where, and by how much, this person's brain departs from the healthy version of itself, rather than from a population-average template. We then use healing fields as the quantitative trait in genetic association testing. The central test is a dissociation. If healing fields isolate disease, they should recover more known Alzheimer's risk loci than template-based measures do, and recover fewer loci that govern normal brain morphology.

Core Claim

The claim has two parts. Measuring each subject's deviation against their own nearest healthy brain removes heritable normal-anatomy variance that template-relative and atlas-based traits keep. Writing that deviation as a deformation, not as an intensity difference, makes reconstruction-based anomaly detection usable for diffuse atrophy.

The second part matters because reconstruction methods are known to handle diffuse atrophy poorly. They detect bright, focal lesions well but struggle with "subtle, diffuse structural deformations, such as tissue atrophy or ventricular enlargement" (DeScarGAN). Atrophy is mainly a change of shape, not a new bright spot, so intensity residuals are a poor fit for it. DeScarGAN itself showed that letting the generator change shape helps on synthetic shrinkage. It reports a Dice score of 0.853 against 0.766 for FP-GAN, but it needs healthy-versus-diseased labels to train. We keep the shape-changing idea and drop the labels.

How the Healing Field Is Computed

First, a healthy prior $p_H(x \mid c)$ is learned. It is a density over T1-weighted MRI volumes $x$, given covariates $c$ (age, sex, intracranial volume), and it is trained only on cognitively unimpaired, amyloid-negative ADNI participants plus radiologically normal UK Biobank scans. The prior can be a three-dimensional variational autoencoder, which is a network that compresses a scan to a small code and decodes it back. It can also be a latent diffusion model, which generates samples by gradually removing noise. We will compare both.

For a new subject, we then search for a smooth, invertible warp $\phi_v$, defined by a velocity field $v$, that makes the warped scan look healthy while moving it as little as possible:

$$ v^\ast = \arg\min_v \; -\log p_H(x \circ \phi_v \mid c) + \lambda \lVert \nabla v \rVert^2 . $$

Here $x \circ \phi_v$ is the scan resampled through the warp, and $\lambda$ sets how much deformation is allowed. The healing field is the log-Jacobian determinant of $\phi_{v^\ast}$, the local fraction of volume that must be restored at each voxel. A small amortized network is trained to predict $v^\ast$ in a single forward pass. This avoids the cost of iterative 3D diffusion sampling, which the diffusion-in-MRI survey flags as a practical barrier (A Survey of Emerging Applications of Diffusion Probabilistic Models in MRI).

Two design choices target known failure modes. The first is the "identical shortcut," in which a flexible model reproduces the anomaly and the residual vanishes. Our objective guards against it because a warp can only make an atrophied brain look healthy by expanding tissue, and the smoothness penalty prevents the model from simply copying the input. The second is leakage. A subject who was in the healthy training set would trivially get a near-zero field and bias the genetic analysis. We therefore cross-fit: every subject's field comes from a model trained on folds that exclude that subject.

Relation to Prior Work

Two genetics papers motivate this proposal and constrain it.

TransferGWAS feeds VAE and ImageNet features through global principal component analysis into a genome-wide association study (GWAS), finding 289 loci in 36,311 UK Biobank scans. Its encoder was not restricted to healthy brains, it builds no counterfactual, and its global components discard location. Its many loci are mostly normal-morphology loci, which makes them a natural negative-control set for us.

Kim and colleagues' spatial enrichment (Identifying genetic markers enriched by brain imaging endophenotypes) correlates each SNP's effect pattern across 116 regions with the case-versus-control disease pattern. A SNP, or single-nucleotide polymorphism, is a single-letter genetic variant. This is the strongest existing method on our question, and we reuse its scoring step unchanged. What we replace is the trait it scores: atlas-region averages of raw imaging measures become voxel-level healing fields.

Normative-modeling papers supply the healthy-reference idea, but none of them test genetics. mmSIVAE builds a healthy reference distribution over 90 FreeSurfer regions using an introspective VAE and scores subjects by Mahalanobis distance. The Focal Adversarial Autoencoder does the same for connectivity matrices and is judged by diagnostic sensitivity, 72.20% on ADNI. Both work on tabular region summaries and produce no voxel maps.

The anomaly-detection papers supply machinery. The VQ-VAE transformer approach (Unsupervised Brain Anomaly Detection and Segmentation with Transformers) "heals" UK Biobank scans by resampling unlikely latent tokens, but it targets focal hyperintensities and computes intensity residuals. We use it as the intensity-residual baseline. Diffusion anomaly localization (Denoising Diffusion Models for Anomaly Localization) documents the noise-depth trade-off: noising too little keeps the anomaly, and noising too much erases the subject's own anatomy. Constraining healing to a regularized warp is our answer to that trade-off, and $\lambda$ plays the role their noise depth plays.

Research Program

The work has three stages.

The first stage builds and validates the healing model before any genetics. Healthy held-out subjects should produce fields indistinguishable from noise. We also implant synthetic atrophy into healthy UK Biobank scans, warping chosen subregions by known amounts, and measure whether the healing field recovers the implanted Jacobian, compared against template-based tensor-based morphometry.

The second stage runs a simulated genetic test with known ground truth. We take real UK Biobank genotypes, choose a SNP, and add atrophy whose size depends on genotype to the scans. One version places the effect inside an atlas region, and another places it across region boundaries. We then measure the power of each trait to detect the planted SNP. This directly tests the region-of-interest bias that both TransferGWAS and Kim and colleagues acknowledge.

The third stage is the real analysis in ADNI, using genotyped participants with T1 MRI, comparable to the 971-subject cohort of Kim and colleagues. Here we run voxelwise and enrichment analyses on healing fields and on every baseline, and score each against the 54 established loci Kim and colleagues used. A secondary arm repeats the analysis on AV-45 amyloid PET, using an intensity-healing variant, because Kim and colleagues' headline number comes from that modality.

Baselines

The naive reference point is a standard case-control GWAS on diagnosis in the same ADNI subjects. Kim and colleagues report that it recovers "2" of the 54 known loci (Identifying genetic markers enriched by brain imaging endophenotypes). We rerun it rather than quote it.

The strongest existing method is that paper's spatial enrichment: per-SNP regression on 116 regions, correlated with the disease map. It reports "41 out of 54" known loci on AV-45 PET in 971 ADNI subjects. No reported result was found on Lacuna for the same pipeline applied to T1 MRI, so we reimplement it on our exact T1 and PET subjects, covariates, and genotype quality control. Healing fields are then scored by the identical enrichment statistic, either pooled to the same 116 regions or kept at voxel level with permutation-calibrated significance. Beating this baseline means recovering more than 41 of 54 loci on PET, or more than our reimplemented T1 figure, at a matched false-positive rate set by permuting genotypes.

The baseline we consider the real threat is tensor-based morphometry. This method warps each brain to a population template and uses the log-Jacobian of that warp as the trait. It differs from our method in exactly one respect: the reference is the population template instead of the subject's healed self. It will use the same registration network, the same smoothness penalty, the same enrichment scoring, and the same tuning budget.

The remaining comparisons each isolate one ingredient. Region-level normative z-scores in the style of mmSIVAE, fed into the same GWAS, test whether voxel-level resolution matters. TransferGWAS-style latent principal components, with the encoder retrained on healthy brains only, test whether an explicit counterfactual matters. VQ-VAE transformer intensity residuals test whether a deformation is needed instead of an intensity difference.

Finally, two ablations isolate the key ingredients. The first sweeps $\lambda$, which moves the method between our approach and template registration. The second drops cross-fitting, to measure leakage.

Why not just…?

Why not just use tensor-based morphometry against a template? This is the closest and most serious alternative, because it already produces voxel-level deformation traits without any generative model. Our argument is that template Jacobians carry heritable normal-shape variance, which healing fields subtract out. That argument is an empirical bet. The dissociation experiment settles it. Healing fields should beat template Jacobians on recovery of the 54 known loci and simultaneously show weaker association with the TransferGWAS normal-morphology loci. They should also show lower SNP-based heritability in healthy UK Biobank participants. If template Jacobians match healing fields on both counts, the counterfactual adds nothing, and we will report that.

Why not just apply normative z-scores to FreeSurfer regions, as mmSIVAE does? Region averages cannot see sub-regional or cross-boundary effects. The cross-boundary simulation in stage two measures that loss directly. If the region-level trait performs as well as the voxel-level trait on the real ADNI loci, the voxel-level machinery is not justified for this cohort size.

Why not just run GWAS on TransferGWAS deep features? The 289 loci from those features show that deep features are strongly heritable. The features are not trained to exclude normal anatomy, however, and they are not spatially localized. Our prediction is that they will be rich in morphology loci and poor in AD-specific enrichment. The healthy-encoder variant tests whether training on healthy brains alone, without healing, already gets most of the benefit.

Why not just use intensity residuals from existing healers? The DeScarGAN paper and the transformer paper both indicate that intensity residuals are suited to focal lesions, not atrophy. The synthetic-atrophy stage quantifies this shortfall before any genetic comparison.

Decisive Checks

The proposal succeeds if three things hold together. Healing fields must recover the implanted Jacobian in synthetic data better than template morphometry. They must give higher power for planted SNP effects, especially effects that cross region boundaries. And in ADNI they must recover more of the 54 known loci than reimplemented spatial enrichment and template morphometry, while carrying less normal-morphology genetic signal than either. Gaining Alzheimer's loci without losing morphology loci would mean the healing step only adds resolution, not disease specificity, and would weaken the core claim.

Cohort size limits what we can conclude. With roughly a thousand ADNI subjects, the evaluation relies on recovery of known loci rather than new discovery. Any new candidate locus is reported only if its spatial pattern replicates on UK Biobank healing fields, which also probes generalization across scanners.